Digital Twins & Synthetic Control Arms

A digital twin in the clinical trial context is an AI-generated, patient-specific prediction of how an individual trial participant would have progressed under a reference condition (typically standard of care or placebo), built from that participant’s own baseline characteristics combined with statistical patterns learned from historical trial data, registries, and other real-world data. The goal is not to replace a control group with pure simulation, but to use these individualized forecasts to make a real, smaller control arm carry more statistical weight — reducing the number of trial participants who need to be randomized to placebo or standard-of-care while preserving the study’s statistical power.

This is an actively developing area of methodology and regulatory policy, not a settled one. This page describes the concept, how it is built, the strongest verifiable regulatory precedent to date, and where the open questions and skepticism sit as of mid-2026. Claims about specific regulatory status are marked as such and should be checked against primary sources before being relied on for a live protocol, since guidance in this area is genuinely still moving.

What “digital twin” means in this context — and what it doesn’t

The term is used loosely across the industry, so it is worth separating three related but distinct ideas that are often conflated:

AI-generated digital twin / prognostic covariate adjustment. A machine-learning model, trained on data from prior trials and/or real-world data sources, generates a predicted outcome trajectory for each newly enrolled participant based on their baseline covariates. That prediction is used as a statistical covariate to sharpen the treatment-effect estimate from a real (not simulated) randomized control arm — it augments the control arm’s statistical efficiency rather than removing the control arm itself. This is the approach most directly associated with the term “digital twin” in current trial methodology discussion.
Synthetic or external control arm. A longer-standing, distinct approach in which some or all of a trial’s control-group comparison is drawn from existing real-world data or prior trial data (e.g., a historical cohort, a registry, or a previous trial’s placebo arm) rather than from newly randomized participants, most often used in single-arm trials for rare diseases or oncology where randomizing a concurrent control is difficult or unethical. FDA has engaged with external control arm methodology for longer than it has engaged with AI-generated digital twins specifically, and the two are sometimes discussed together but are not the same technique.
Full patient simulation replacing a control arm entirely. This — a trial with no real control participants at all — is not current regulatory practice for pivotal trials. Every publicly documented precedent to date uses digital twins to augment or shrink a real control arm through statistical covariate adjustment, not to eliminate it.

How a digital twin / prognostic covariate adjustment model is built

The general pipeline described across vendors and published methodology work follows a consistent pattern:

Training data assembly. Historical clinical trial datasets (often from completed trials in the same disease area, sometimes licensed from prior sponsors or aggregated by a vendor across many past studies) and, in some approaches, real-world data such as registries or electronic health records.
Disease progression modeling. A machine-learning model is trained to predict how a patient’s disease course and trial endpoints evolve over time, conditioned on baseline characteristics (demographics, disease severity, biomarkers, comorbidities, and other covariates collected at enrollment).
Per-participant prediction at enrollment. When a new participant enrolls in the live trial, the trained model generates a predicted outcome trajectory for that specific participant under the reference (control) condition, using only their baseline data — collected before randomization, so the prediction cannot be influenced by which arm they are later assigned to.
Statistical use in the analysis. The prediction is incorporated into the trial’s pre-specified statistical analysis as a covariate adjustment, which reduces the variance of the estimated treatment effect. Lower variance means the trial can detect the same effect size with fewer randomized control-arm participants, or detect a smaller effect size with the same sample — the mechanism by which “digital twins” translate into a smaller placebo/control group rather than a total absence of one.

Because the prediction is generated from pre-randomization baseline data only, and is used as a covariate rather than as a substitute outcome, this design differs from naively “simulating” trial results — the actual outcome data from real randomized participants still drives the primary treatment-effect estimate.

The clearest regulatory precedent: Unlearn.AI and PROCOVA

The most frequently cited real-world precedent is Unlearn.AI’s PROCOVA (Prognostic Covariate Adjustment) methodology, which is publicly reported to have received a qualification opinion from the European Medicines Agency (EMA) in September 2022 — the first time a European regulator formally qualified a machine-learning-based covariate adjustment method for use in reducing sample size in pivotal trials. Unlearn.AI has also reported receiving positive FDA feedback supporting use of the method in covariate-adjusted analyses. This is REPORTED-tier information corroborated across the vendor’s own published materials and independent industry coverage rather than a primary EMA/FDA document reviewed directly for this page — sponsors evaluating the method for a live protocol should confirm current qualification status and scope directly with EMA and FDA rather than relying on secondary summaries, since qualification opinions are typically scoped to a specific methodology and context of use, not a blanket endorsement of “digital twins” as a category.

FDA’s regulatory posture: engaged, but explicitly cautious and not yet settled

It is important not to overstate where FDA currently stands. The clearest primary-source evidence of FDA’s posture toward AI-generated evidence generally (not digital twins specifically) is a draft guidance, “Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products,” issued January 6, 2025 (comment period closed April 7, 2025; as of this writing it remains draft guidance, not final). The draft guidance proposes a risk-based, seven-step credibility assessment framework for any AI model used to support a regulatory submission: define the question of interest, define the context of use, assess model risk, develop and execute a credibility assessment plan, document results and any deviations, and determine whether the model’s output is adequate for its specific context of use. This framework applies broadly to AI use across nonclinical, clinical, post-marketing, and manufacturing phases — it is not a digital-twin-specific policy, but it is the framework a sponsor proposing a digital-twin-augmented control arm would need to work within.

Separately, FDA’s longer-running Complex Innovative Trial Design (CID) pilot meeting program (established under PDUFA VI) gives sponsors a channel to discuss novel statistical and trial-design approaches — including model-informed and simulation-based designs — with the agency before a pivotal trial, and digital-twin-augmented designs are the kind of approach sponsors have used that channel for. A CID pilot meeting is a discussion mechanism, not an approval or a blanket endorsement, and outcomes are specific to the sponsor and protocol involved.

Taken together: FDA has engaged directly with AI-generated covariate adjustment and has given positive feedback in at least the Unlearn.AI/PROCOVA case reported above, and it has published a general framework for assessing AI model credibility. That is meaningfully different from FDA having issued a general endorsement, qualification, or approval of “digital twins” as a category for pivotal trial control arms — each use case is still evaluated on its own methodology, data provenance, and context of use. Sponsors should treat this as an active, evolving regulatory conversation and confirm current agency posture directly (including through FDA’s own guidance page and, where appropriate, a CID or other pre-submission meeting) rather than assuming a settled pathway.

Open questions and the validation debate

Coverage of digital twins in trial design consistently raises a set of unresolved concerns that sponsors and IRBs are actively working through, rather than issues that have been fully resolved:

External validity of the training data. A digital twin model is only as good as the historical trials and real-world data it was trained on — if that training population differs meaningfully from the population being enrolled in the new trial (in demographics, standard of care, disease definitions, or era of data collection), the model’s predictions may not transfer well.
Model drift and generalizability across indications. A model validated for one disease area or endpoint does not automatically generalize to another; each context of use is expected to require its own credibility assessment under frameworks like FDA’s.
Transparency and reproducibility. Because these are proprietary machine-learning models built on licensed or aggregated datasets, independent replication of a specific vendor’s model is harder than replicating a conventional statistical method described in a published protocol.
Where the line sits between “smaller control arm” and “no real control arm.” The current precedent is variance reduction and sample-size reduction on a real control arm, not elimination of a concurrent control group — but as the methodology matures, this is exactly the boundary regulators, methodologists, and ethicists are debating.

Why sponsors are pursuing this

The underlying motivation is consistent across the reporting on this topic: fewer participants need to be randomized to placebo or standard-of-care, which can reduce enrollment timelines, lower the ethical burden of withholding an investigational treatment from participants (particularly relevant in serious or rare diseases with few effective alternatives), and reduce overall trial cost — all without changing what the trial is powered to detect, if the covariate adjustment is done well. This is the same underlying motivation behind the broader push toward decentralized clinical trial designs and adaptive methodologies more generally: using statistical and operational innovation to make a trial less burdensome to run without weakening its evidentiary value.

How this relates to other AI use in trials and to data monitoring

Digital twins are one specific, statistically-grounded application within the broader landscape covered in CASRAI’s AI in Clinical Trials guide — which surveys patient recruitment matching, site feasibility, protocol simulation, and safety-signal detection more broadly and situates this same FDA draft guidance in that wider context. For the general regulatory framework FDA is building for AI across drug development and device review, see CASRAI’s FDA AI Guidance guide.

Because a digital-twin-augmented control arm still relies on a real, monitored control group, the same data-quality and oversight expectations that apply to any trial arm still apply here — see CASRAI’s dictionary entry on Risk-Based Monitoring (RBM) for how sponsors prioritize monitoring resources toward the data points that matter most to a trial’s conclusions, and the Clinical Trial Monitoring guide for how monitoring visit types and source data verification work in practice.

Frequently asked questions

Is a digital twin the same as a synthetic control arm?

Not exactly. “Synthetic control arm” and “external control arm” more commonly refer to using an existing dataset (a prior trial’s control arm, a registry, or real-world data) as some or all of a trial’s comparator, a technique FDA has engaged with for longer, often in single-arm oncology or rare-disease trials. “Digital twin” as currently used most often refers specifically to an AI model generating a per-participant predicted outcome used as a statistical covariate to shrink — not replace — a real, concurrently randomized control arm. The terms are sometimes used loosely and interchangeably in industry writing, so it is worth clarifying which specific technique a given source or vendor means.

Has the FDA approved digital twins for clinical trials?

Not as a blanket policy. FDA has given positive feedback on at least one specific methodology (Unlearn.AI’s PROCOVA, as publicly reported) and has published a general draft framework (January 2025) for assessing the credibility of any AI model used in a regulatory submission, but there is no general FDA approval or qualification of “digital twins” as a category. Each proposed use is evaluated on its own data, model, and context of use, and the January 2025 framework remains in draft form as of this writing.

Does using a digital twin mean fewer patients get the investigational treatment?

No — the effect runs the other way. Because the technique reduces the number of participants needed in the control/placebo arm while preserving statistical power, a smaller share of total enrolled participants receive placebo or standard-of-care alone, and a larger relative share can be randomized to (or offered access to) the investigational arm, for the same overall trial size.

What data is a digital twin model trained on?

Historical clinical trial datasets in the same or a related disease area, and in some approaches real-world data such as patient registries or electronic health records, used to train a model that predicts individual disease-progression trajectories from baseline covariates collected before randomization.

Digital Twins in Clinical Trials: Synthetic Control Arms and FDA’s Evolving Posture